Xiaomi launches three MiMo AI models to power agents, robots and voice assistants

Xiaomi’s three‑model MiMo suite marks the company’s first full‑stack offering for autonomous AI agents, a move that could reshape the competitive landscape for Chinese‑origin large‑language models (LLMs). The flagship MiMo‑V2‑Pro, built on a mixture‑of‑experts (MoE) architecture with more than one trillion parameters—though only 42 billion are active per request—delivers performance that rivals Anthropic’s Claude Opus 4.6 on both coding and agent benchmarks while costing a fraction of the price. According to The‑Decoder, MiMo‑V2‑Pro ranks third globally on PinchBench and ClawEval, trailing only Claude Opus 4.6, and lands seventh on the Artificial Analysis Intelligence Index, making it the top‑performing Chinese model after GLM‑5 and MiniMax‑M2.7. On the SWE‑bench Verified coding benchmark it scores 78 percent, just shy of Claude Opus 4.6’s 80.8 percent, and on the ClawEval agent benchmark it posts 81 points, essentially matching Claude Opus 4.6’s 81.5.

Pricing is a central pillar of Xiaomi’s strategy. The company lists MiMo‑V2‑Pro at US$1 per million input tokens and US$3 per million output tokens for context windows up to 256 k tokens, with all cache‑writing fees waived for the launch period. By contrast, Claude Sonnet 4.6 charges US$3–15 per million tokens and Claude Opus 4.6 costs US$5–25, according to the same source. The lower cost structure, combined with a public API and a one‑week free‑access window for developers, is intended to accelerate adoption across the five agent frameworks Xiaomi has partnered with—OpenClaw, OpenCode, KiloCode, Blackbox and Cline.

The multimodal MiMo‑V2‑Omni model extends the platform’s reach beyond text. It integrates image, video, and audio encoders into a shared backbone, enabling the model to “see, hear, and act” autonomously. The Decoder reports that MiMo‑V2‑Omni beats Claude Opus 4.6 on audio and image benchmarks, though it lags behind Google’s Gemini 3 Pro on video tasks. The model supports structured tool calls and can navigate user interfaces without external orchestration, a capability Xiaomi highlights for use cases such as browser‑based shopping or real‑time dash‑cam hazard analysis. The company also claims continuous audio recording for over ten hours, a metric that could be valuable for long‑form voice assistants and robot perception.

Xiaomi’s broader AI ambitions are underscored by a parallel announcement from Reuters that the firm will invest at least US$8.7 billion in AI over the next three years. CEO Lei Jun framed the investment as a “strategic push” to build AI‑enabled hardware—from smartphones to household robots—leveraging the new MiMo models as the software core. VentureBeat echoed this narrative, noting that MiMo‑V2‑Pro’s performance approaches that of GPT‑5.2 while remaining far cheaper, positioning Xiaomi as a serious contender not only against Western providers but also against domestic rivals such as Baidu and Alibaba, which have focused more on cloud‑centric offerings.

Analysts will watch how quickly developers adopt the MiMo stack, especially given the free‑week API incentive and the integration with multiple agent frameworks. If the platform can deliver reliable autonomous behavior—e.g., seamless browser navigation, robust multimodal perception, and low‑latency token generation enabled by its hybrid attention mechanism and parallel token output—Xiaomi could capture a sizable share of the burgeoning enterprise‑agent market. The company’s ability to scale the MoE architecture efficiently, maintain token‑cost advantages, and translate benchmark success into real‑world productivity will determine whether MiMo becomes a cornerstone of China’s AI ecosystem or remains a high‑profile but niche offering.